Electronic Notes in Theoretical Computer Science ?? (2004) URL: http://www.elsevier.nl/locate/entcs/volume??.html ?? pages
Semantic Domains for Combining Probability and Non-Determinism
Regina Tix Klaus Keimel and Gordon Plotkin
Fachbereich Mathematik, Technische Universit¨atDarmstadt, Germany University of Edinnburgh, UK
c 2004 Published by Elsevier Science B. V.
Contents
Foreword v
Introduction vii
1 Order and Topology 1 1.1 Dcpos and Scott-Continuous Functions ...... 1 1.2 The Specialisation Order ...... 3 1.3 Sober Spaces ...... 5 1.4 Continuous Domains ...... 5 1.5 Lawson-Compact Continuous Domains ...... 7
2 Directed Complete Ordered Cones 11 2.1 D-Cones and Their Basic Properties ...... 12 2.1.1 The Way-Below Relation ...... 15 2.1.2 Convex Sets ...... 16 2.2 The Extended Probabilistic Powerdomain ...... 19 2.2.1 Valuations and Measures ...... 25 2.2.2 Additivity of the Way-Below Relation on the Extended Probabilistic Powerdomain ...... 29 2.3 Lower Semicontinuous Functions and Dual Cones ...... 32
3 Hahn-Banach Type Theorems 37 3.1 A Sandwich Theorem ...... 38 3.2 A Separation Theorem ...... 42 3.3 A Strict Separation Theorem ...... 45 3.4 An Extension Theorem ...... 48
iii iv
4 Power Constructions 55 4.1 The Convex Lower Powercone ...... 57 4.1.1 The Convex Lower Powercone Construction ...... 57 4.1.2 Universal Property of the Convex Lower Powercone . 62 4.2 The Convex Upper Powercone ...... 65 4.2.1 The Convex Upper Powercone Construction ...... 66 4.2.2 Universal Property of the Convex Upper Powercone . 71 4.3 The Biconvex Powercone ...... 76 4.3.1 The Biconvex Powercone Construction ...... 77 4.3.2 Universal Property of the Biconvex Powercone . . . . 85 4.4 Powerdomains Combining Probabilistic Choice and Non-Determinism ...... 89
Conclusion: Some Connections with Semantics 93
Bibliography 101
Index 106 Foreword
This volume is based on Regina Tix’s 1999 doctoral dissertation [55], entitled Continuous D-cones: Convexity and Powerdomain Constructions and sub- mitted to the Department of Mathematics of Technische Universit¨atDarm- stadt. Only a small part of this thesis, namely three sections of Chapter 3, has previously been published (see [56]). Since then, the main body of the thesis, Chapter 4 on powerdomains for modelling non-determinism, has be- come of increasing interest: indeed the main goal of the thesis was to provide semantic domains for modelling the simultaneous occurrence of probabilistic and ordinary non-determinism. It therefore seemed appropriate to make the thesis available to a general audience. There has been a good deal of progress in the relevant domain theory since the thesis was submitted, and so Klaus Keimel has rewritten large parts of the text, while maintaining the global structure of the original dissertation. As well as making a great number of minor changes, he has incorporated some major improvements. Gordon Plotkin has proved a Strict Separation Theorem for compact sets: all of Section 3.3 is new and essentially due to him. The Strict Sep- aration Theorem 3.8 enables us, in Chapter 4, to eliminate an annoying auxiliary construction used in the original thesis for both the convex upper and the biconvex powercones; one also gets rid of the requirement that the way-below relation is additive, and the whole presentation becomes simpli- fied and shorter. Next, an annoying hypothesis of a non-equational nature is no longer required for the statement of the universal property of the biconvex pow- ercone. Further, the hypotheses for the lower powercone have been weak-
v vi FOREWORD ened: the universal property for this powercone remains valid without re- quiring the base domain to be continuous. Finally, we have added Section 4.4 explicitly presenting the powerdomains combining probabilistic choice and non-determinism and their universal properties. Combining the extended probabilistic powerdomain with the classical convex powerdomain was not possible when Tix’s thesis was submitted: it was not known then whether the extended probabilistic powerdomain over a Lawson-compact continuous domain is Lawson-compact. Extending slightly a recent result from [3], we now know that the extended probabilistic powerdomain is Lawson-compact over any stably locally compact space. For continuous domains the converse also holds. This allows us in particular to include infinite discrete spaces. We have included these new results in section 2.2. There have also been some terminological changes. For the classical powerdomains we now speak of the lower, upper, and convex powerdo- mains instead of the Hoare, Smyth, and Plotkin ones. Accordingly, for the new powerdomains we speak of the convex lower, convex upper, and bicon- vex powercones, rather than the convex Hoare, convex Smyth, and convex Plotkin powercones. D. Varacca [57, 58, 59] took a related approach to combining probability and nondeterminism via indexed valuations. His equational theory is weaker; he weakens one natural equation, but the theory becomes more flexible. M. Mislove [37] has introduced an approach similar to ours for the probabilistic (not the extended probabilistic) powerdomain, his goal being a semantics for probabilistic CSP. It is quite likely that our results can be used to deduce analogous properties for the (restricted) probabilistic powerdomain. Without the 2003 Barbados Bellairs Workshop on Domain Theoretic Methods in Probabilistic Processes and the inspiring discussions there, in particular with Franck van Breugel, Vincent Danos, Jos´ee Deharnais, Mart´ın Escard´o,Achim Jung, Michael Mislove, Prakash Panangaden, and Ben Wor- rell, this work would not have been undertaken. Achim Jung’s advice has been most helpful during the preparation of the manuscript. The diagrams were drawn using Paul Taylor’s diagrams macro package.
December 2004 Regina Tix, Klaus Keimel, Gordon Plotkin Introduction
The semantics of programming languages has been intensively studied by mathematicians and computer scientists. In the late sixties Dana S. Scott invented appropriate semantic domains for that purpose [51, 49, 50]. Con- tinuous domains are directed complete partially ordered sets together with an order of approximation, the so called way-below relation. As they al- low one to represent ‘ideal objects’ and their ‘finite approximations’ within one framework, continuous domains provide a suitable universe for denota- tional semantics. The order can be thought of as an ‘information ordering’. That means the greater an element the more information it carries about the object it approximates. In this approach, computable functions are con- tinuous functions on domains. Moreover, within domains, recursion can be interpreted via least fixed points of continuous functions. Domain theory has since attracted many researchers and evolved in various directions. It owes much to the theory of continuous lattices and domains, most notably [14, 15]. An important problem in domain theory is the modelling of non-determi- nistic features of programming languages and of parallel features treated in a non-deterministic way. If a non-deterministic program runs several times with the same input, it may produce different outputs. To describe this behaviour, powerdomains were introduced by Plotkin [40, 41] and Smyth [52]. A powerdomain over a domain X is a subset of the power set of X. Which subsets of X constitute the powerdomain depends on the kind of non-determinism that is be modelled. There are three classical power- domain constructions, called the convex, upper, and lower powerdomains, often referred to as Plotkin, Smyth, and Hoare powerdomains.
vii viii INTRODUCTION
Probabilistic non-determinism has also been studied and has led to the probabilistic powerdomain as a model [47, 42, 24, 23]. Different runs of a probabilistic program with the same input may again result in different outputs. In this situation, it is also known how likely these outputs are. Thus, a probability distribution or continuous valuation on the domain of final states is chosen to describe such a behaviour. Originally attention had been paid to valuations with total mass ≤ 1. This leads to powerdomains carrying a convex structure. The collection of all continuous valuations (bounded or not) on a continuous domain X, ordered ‘pointwise’, leads to the extended probabilistic powerdomain of X. The extended probabilistic powerdomain carries the structure of a cone, more technically of a continuous d-cone [29], a structure close to that of a an ordered cone in a topological vector space as considered in functional analysis. This development led to an intrinsic interest in d-cones (see also Chapter 2). For Plotkin’s and Jones’ model of probabilistic computation the continu- ous d-cone of lower semicontinuous, i.e., Scott-continuous, functions defined on the domain X with values in the non-negative extended reals is also needed. Integration of such lower semicontinuous functions with respect to a continuous valuation plays a crucial role. One obtains a duality between the extended probabilistic powerdomain over a continuous domain X and the continuous d-cone of lower semicontinuous functions on X. One direction of this duality is given by a version of the Riesz’ Representation Theorem. This leads to functional analytic questions about continuous d-cones and their duals for example: whether there exist non-zero linear Scott-continuous functionals, and whether these separate points. We will discuss this issue among other Hahn-Banach type theorems in Chapter 3. It still is an open problem whether there is a cartesian closed category of continuous domains which is closed under the construction of probabilistic powerdomains. This issue is discussed in [25]. Cartesian closure is essential in the denotational semantics of functional languages. There is a new challenge: What happens if non-deterministic choice co- exists with probabilistic choice? And how can the classical powerdomain constructions together with the probabilistic powerdomain be used for mod- elling such situations? The Programming Research Group in Oxford [43] has INTRODUCTION ix tackled various aspects of this problem. Out of this group, McIver and Mor- gan have chosen a subdomain of the Plotkin powerdomain over the space of subprobability distributions on discrete state spaces [36]. The subsets they allow are the convex ones. Our approach to convex powercones was motivated by theirs. We modify and generalize their construction to con- tinuous Lawson-compact d-cones. Therefore, we introduce and investigate a Hoare and Smyth style powerdomain construction for continuous d-cones. Then the convex Plotkin powercone can be defined as a combination of the other two constructions. It is our goal to apply these constructions to the extended probabilistic powerdomain in Section 4.4. More background information will be given in the introductory part of each chapter. The course of the work is as follows: Chapter 1 introduces briefly the prerequisites from domain theory used in this work and it recalls the interplay between order and topology in domain theory. Continuous d-cones are the focus of Chapter 2. These are continuous domains which carry the structure of a cone in such a way that addition and scalar multiplication are Scott-continuous. The following examples of continuous d-cones will be investigated: the non-negative extended real num- bers, the extended probabilistic powerdomain over a continuous domain, the cone of lower semicontinuous functions on a core compact space with val- ues in the non-negative extended real numbers, and products of continuous d-cones. We will see that continuous d-cones are always locally convex, in the sense that each point has a neighbourhood basis of Scott-open convex sets (the notion of convexity is that of convex sets in real vector spaces and has to be distinguished from order-convexity). Sometimes, the hypothesis of an additive way-below relation is useful. We will show that this property is satisfied in all of the above examples with one restriction: The d-cone of lower semicontinuous functions has an additive way-below relation if and only if the underlying space is coherent. We will also give a brief exposition on the relation between continuous valuations and Borel measures. In Chapter 3, Hahn-Banach type theorems for continuous d-cones will be proved. We begin by proving a Sandwich Theorem. From this we obtain Separation Theorems. Since continuous d-cones are locally convex, the Sepa- x INTRODUCTION ration Theorems imply that the Scott-continuous linear functionals separate the points on a continuous d-cone. The Strict Separation Theorems will be needed for the convex upper and biconvex powercones. Another application of the Separation Theorem will be indicated in the Conclusion: in connec- tion with semantics it can be used to show that a special map between two models is injective. Extension Theorems are another type of Hahn-Banach Theorems. We will prove a typical extension theorem for continuous d-cones with an additive way-below relation. Chapter 4 introduces Hoare, Smyth and Plotkin style constructions for continuous d-cones with the intention to apply them to the extended prob- abilistic powerdomain. However, the constructions are feasable and more transparent in the general setting of continuous d-cones. First, we modify the topological characterisation of the lower powerdomain by taking only those non-empty Scott-closed subsets which are also convex. This allows us to lift addition and scalar multiplication in such a way that we obtain a d- cone again, called the convex lower powercone. In addition, binary suprema exist in the convex lower powercone and the convex lower powercone is shown to be universal in this context. For the upper powerdomain we replace non-empty convex Scott-closed sets by non-empty convex compact saturated sets. Again, this enables us to lift the algebraic operations. We also obtain a d-cone, this one with binary infima as extra semilattice operation. However, for this d-cone continuity is equivalent to the existence of linear Scott-continuous functionals which separate compact saturated convex sets from points. The convex upper construction is universal in a suitable setting with respect to binary infima. The biconvex powercone can be defined over Lawson-compact continuous d-cones as a combination of the convex lower powercone and the convex upper powercone. We prove that the biconvex powercone is also Lawson- compact, and that it is universal in this setting with respect to a binary semilattice operation, called formal union. This work concludes with giving an idea on how its results can be used for semantics in a situation, where non-deterministic features can be denoted alongside probabilistic ones. Chapter 1
Order and Topology
In this chapter we briefly review the prerequisites on order and topology necessary for our further results. The focus will be on domain theory; how- ever, a complete introduction to this topic by far exceeds the scope of this work. Thus, we present selected items only and omit all proofs as we go along to fix our notation. We refer to [1, 7, 14, 15, 33] for more details.
1.1 Dcpos and Scott-Continuous Functions
We shall use the term ordered set in the sense of partially ordered set, that is, it denotes a set X with a reflexive, antisymmetric and transitive binary relation ≤, not necessarily linear. For any subset A of X we get the lower, resp. upper, closure of A by
↓A := {x ∈ X | x ≤ a for some a ∈ A} , ↑A := {x ∈ X | x ≥ a for some a ∈ A} .
We abbreviate ↓{a} to ↓a and ↑{a} to ↑a. A subset A with A = ↓A is called a lower set; A = ↑A is called an upper set. A subset D of an ordered set X will be called directed if it is nonempty and if any two elements of D have a common upper bound in D. The dual notion is that of a filtered set. An ordered set X will be called directed complete or a dcpo, for short, if each directed subset D has a least upper ↑ bound W D in X. If this is true only for directed subsets that are bounded
1 2 CHAPTER 1. ORDER AND TOPOLOGY from above, then we say that X is conditionally directed complete. If every subset A has a least upper bound sup A = W A, then X is a complete lattice. The least upper bound of any (directed) subset is also called its (directed) supremum. The set R+ of non-negative real numbers with the usual total order is conditionally directed complete, whilst R+ = R+ ∩{+∞} is directed complete. A function f : X → Y between ordered sets is order preserving or mono- tone, if a ≤ b implies f(a) ≤ f(b) for all a, b ∈ X. If X and Y are (condi- tionally) directed complete, then f : X → Y is called Scott-continuous, if it ↑ ↑ is order preserving and if f(W D) = W f(D) for every (bounded) directed subset D ⊆ X. When we talk about continuous functions between (condi- tionally) directed complete partial orders, we always mean Scott-continuous functions. We denote by DCPO the category of dcpos and Scott-continuous functions. The least upper bound of a directed set D may be considered as a limit of D. This explains the choice of the notion of continuity. This can be made precise with respect to an appropriate topology: A subset A of a (conditionally) directed complete ordered set X will be called Scott-closed ↑ if A is a lower set and if W D ∈ A for every (bounded) directed set D ⊆ A. The complement X \ A of a Scott-closed set A will be called Scott-open. Thus, a set U is Scott-open, if U is an upper set and if for every (bounded) ↑ directed subset D of X the following holds: If W D ∈ U, then d ∈ U for some d ∈ D. It is easily seen that the Scott-open sets form a topology on X, the Scott topology. This topology always fulfills the T0-separation axiom, but is non-Hausdorff unless the (conditionally) directed complete partial order is ordered trivially. Throughout this work, A will denote the closure of a subset A of a (conditionally) directed complete partial order with repect to the Scott topology. Scott continuity as defined above is consistent with the Scott topology: A function f : X → Y between (conditionally) directed complete ordered sets is Scott-continuous if and only if f is continuous with respect to the Scott topologies on X and Y . A product X × Y of (conditionally) directed complete ordered sets X and Y is again (conditionally) directed complete. A function f defined 1.2. THE SPECIALISATION ORDER 3 on X × Y is Scott-continuous if, and only if, it is componentwise Scott- continuous, that is, if x 7→ f(x, y) is Scott-continuous on X for every fixed y ∈ Y and similarly for the second component. It is an unfortunate fact that the Scott topology on X × Y may be strictly finer than the product of the Scott topologies on X and Y , unless one of X and Y is continuous (see sec. 1.4 and [15, p. 197]). Thus, a Scott-continuous function defined on X × Y need not be continuous for the product topology unless one of X and Y is continuous. For any topological space X we denote the collection of open sets by O(X). Ordered by set inclusion, this gives a complete lattice. Especially, directed suprema exist and O(X) itself can be viewed as a topological space with the Scott topology.
1.2 The Specialisation Order
So far, we have seen how to equip a (conditionally) directed complete ordered set with a T0-topology. Let us now change our point of view and consider a T0-topological space X. Such a space always carries an intrinsic order, namely the specialisation order. It is defined by x ≤ y if x is in the closure of {y} or, equivalently, if the neighbourhood filter of x is contained in the neighbourhood filter of y. This definition always yields a reflexive, transitive relation, which is antisymmetric exactly for T0-spaces. For this reason, a topological space is always supposed to satisfy the T0-separation axiom in this work. In the case of a T1-space, where every singleton set is closed, the specialisation order is trivial. Continuous functions between topological spaces preserve the respective specialisation orders. For the product of topological spaces with the product topology, the specialisation order is equal to the product of the respective specialisation orders. A closed set is always a lower set and every open set is an upper set. The closure of a point is exactly its lower closure {a} = ↓a. Let us recall the following result from [48, Corollary 1.6(i)]:
Lemma 1.1 Let f : X → Y be a continuous map between T0-topological spaces and let A be a subset of X. With respect to the specialisation orders, 4 CHAPTER 1. ORDER AND TOPOLOGY the supremum of f(A) exists in Y if and only if the supremem of f(A) exists in Y . In this case, W f(A) = W f(A).
For a dcpo with the Scott topology the specialisation order coincides with the originally given order. The saturation of any subset A in a topological space is defined to be the intersection of all the neighbourhoods of A. This is exactly the upper closure ↑A with respect to the specialisation order. Thus, an upper set will also be called saturated. In T0-spaces all sets are saturated. It is an immediate consequence of the definition that the saturation of any compact set is again compact. Compactness is defined by the Heine-Borel covering property: every cov- ering by open sets has a finite subcovering. For a monotone map f : X → Y between two ordered sets, in particular, for a continuous map between topo- logical spaces with their specialisation orders, ↑f(↑A) = ↑f(A) holds for any subset A of X. We will mainly apply this to compact saturated subsets and Scott- continuous functions. From general topology we know that the continuity of a function f : X → Z can be characterized by the property that f(A) ⊆ f(A) or, equivalently, f(A) = f(A), for every subset A ⊆ X. We will need the following con- sequence which can be applied to dcpos and Scott-continuous functions on products, which are only separately continuous with respect to the product topology (see the remarks at the end of section 1.1):
Lemma 1.2 Let X,Y,Z be topological spaces and f : X × Y → Z be sepa- rately continuous, that is, x 7→ f(x, y) is continuous on X for every y ∈ Y and similarly for the second coordinate. For all subsets A ⊆ X and B ⊆ Y one then has f(A × B) = f(A × B) = f(A × B).
Proof By separate continuity, we have f(A×{y}) ⊆ f(A × {y}) ⊆ f(A × B) for all y ∈ Y , whence f(A × B) ⊆ f(A × B), and this implies f(A × B) = f(A × B). The second equality follows in an analogous way. 1.3. SOBER SPACES 5
1.3 Sober Spaces
For a special class of T0-spaces every non-empty closed subset is either the closure of a unique point or the union of two proper closed subsets. We call such spaces sober. An equivalent formulation of sobriety is that every completely prime filter of open sets on X is the open neighbourhood filter of a unique point a ∈ X. The collection of all nonempty compact saturated subsets of a topological space is denoted by Sc(X) and will be ordered by reverse inclusion. An important property of sober spaces X is the so called Hofmann-Mislove Theorem (see [20, 26], [15, Theorem II-1.20]). The following proposition (see [15, TheoremII-1.21, Corollary II-1.22]) is a consequence of this theorem. It will be used extensively in Section 4.2.
Proposition 1.3 Let X be a sober space. The intersection of a filtered fam- ily (Qi) of nonempty compact saturated subsets is compact and nonempty. If such a filtered intersection is contained in an open set U, then Qi ⊆ U for some i.
The first part of this proposition can be rephrased as follows: Sc(X) ordered by reverse inclusion is a dcpo for any sober space X. It is another property of sober spaces that the specialisation order yields a dcpo, with the original topology being coarser than the Scott topology. However, a dcpo with the Scott topology is not always sober [22]. In the next section we introduce special dcpos, called continuous domains, which are always sober spaces with respect to the Scott topology [31].
1.4 Continuous Domains
On a (conditionally) directed complete partial order X we introduce a bi- nary relation as follows: Let x and y be elements of X. We say that x approximates y or x is way-below y, and we write x y, if for all (bounded) ↑ directed subsets D of X, the inequality y ≤ W D implies x ≤ d for some d ∈ D. We call the order of approximation or way-below relation on X. It is immediate that x y implies x ≤ y, and w ≤ x y ≤ z implies 6 CHAPTER 1. ORDER AND TOPOLOGY w z, in particular, the way-below relation is transitive. If x ∨ y exists, then x z and y z imply x ∨ y z. For any x ∈ X and for any subset A ⊆ X, we use the notations
x := {y ∈ X | x y} , A := {y ∈ X | x y for some x ∈ A} , x := {y ∈ X | y x} , A := {y ∈ X | y x for some x ∈ A} . A (conditionally) directed complete partial order X is called continuous if, ↑ for all x ∈ X, the set x is directed and x = W x. A continuous dcpo is also called a continuous domain. A subset B of a continuous domain X is called a basis of X if, for all x ∈ X, the set x ∩ B is directed and has x as its supremum. In a continuous domain, a basis always exists, for example take B = X. Moreover, in a continuous domain the so called interpolation property holds: Whenever x y, there is z ∈ X such that x z y. If a basis of X is given, z can be chosen from this basis. We denote the category of continuous domains and Scott-continuous functions by CONT. The Scott topology of a continuous domain can be nicely described via the way-below relation. The sets of the form x, x ∈ X, form a basis of this topology. Again, we can restrict ourselves to a basis B of X, i.e., the sets b, b ∈ B, also form a basis of the Scott topology. The Scott closure of a subset A of an arbitrary dcpo can be obtained in the following way: Let A0 = A and define by transfinite induction An+1 to W↑ be the set of all x such that x ≤ D for some directed subset D of An; S for limit ordinals n, we let An = m Lemma 1.4 In a continuous domain X the Scott closure of an arbitrary subset A is _↑ A = D | D a directed subset of ↓A . For a continuous domain it is known how to obtain the largest Scott- continuous function below a monotone one. The construction once again relies on a monotone function defined on a basis only. 1.5. LAWSON-COMPACT CONTINUOUS DOMAINS 7 Proposition 1.5 Let B be a basis of a continuous domain X and let Y be a dcpo. For every monotone function f : B → Y there is a largest Scott- ˇ ˇ ˇ continuous function f : X → Y such that f|B ≤ f. It is given by f(x) = ↑ W {f(y) | y x and y ∈ B}. Let X and Y be dcpos. Then a pair of Scott-continuous functions r : X → Y and s: Y → X is called a continuous retraction-section-pair if r ◦ s is the identity on Y . Note that in this case r is surjective and s is injective. We will call Y a retract of X, and it can be shown that a retract of a continuous domain is again a continuous domain (see [15], p. 81). We call a space locally compact if every point has a neighbourhood basis of compact sets. Note that continuous domains are always locally compact. Actually, a somewhat stronger condition holds: Lemma 1.6 In a continuous domain each Scott-compact subset has a neigh- bourhood basis of Scott-compact saturated sets. 1.5 Lawson-Compact Continuous Domains According to Nachbin [39], an ordered topological space is a topological space with an order ≤ such that the graph of the order relation is closed in X × X with the product topology. In [14, 15] (partially) ordered topological spaces are called pospaces. One immediately concludes that any pospace is Hausdorff. Another property which can already be found in [39] is the following: Lemma 1.7 Let X be a pospace. If A is a compact subset, then ↓A, ↑A and ↓A ∩ ↑A are closed subsets of X. For any ordered topological space X the collection U(X) of all open upper sets is closed under finite intersections and arbitrary unions, that is, U(X) is a topology on X which is T0 but not Hausdorff unless the order is trivial. Note that the specialisation order with respect to the topology U(X) coincides with the original order on X. On the other hand, given a T0-topological space with its specialisation order, one may define the co-compact topology which has the compact sat- 8 CHAPTER 1. ORDER AND TOPOLOGY urated subsets as a subbasis for the closed sets. The open sets for the co-compact topology are lower sets. The common refinement of a topology with its co-compact topology is called the patch topology. Another way of creating a topology which is coarser than the co-compact topology is by tak- ing as a subbasis of closed sets the principal filters ↑x, x ∈ X. This weakest T0-topology whose open sets are lower sets is called the lower topology. There is an important one-to-one correspondance between compact or- dered spaces and certain classes of T0-spaces to be defined. Definition A topological space X is called coherent, if the intersection of any two compact saturated subsets is compact. It is called stably locally compact, if it is locally compact, sober, and coherent; if, in addition, X is a compact space, then it is called stably compact. Proposition 1.8 ([15, Proposition VI-6.8, Proposition VI-6.11]) Let X be a stably compact space. With respect to the patch topology and the special- isation order, X becomes a compact pospace; the patch-open upper sets are precisely the open sets for the original topology. Conversely, let X be a compact pospace. With respect to the topology U(X) of open upper sets, X becomes a stably compact space the patch topology of which is the original compact topology on X. The corresponding result holds for stably locally compact spaces on the one hand, and properly locally compact pospaces on the other hand, where a pospace is called properly locally compact, if it is locally compact and if ↑K is compact for every compact subset K. A locally compact pospace is far from being properly locally compact, in general; the real line with is usual order and topology, for example, is a non-properly locally compact pospace. We now apply these ideas to dcpos with the Scott topology. For any dcpo the Lawson topology is defined to be the common refinement of the Scott topology and the lower topology. In case the dcpo X is continuous the Scott topology always is locally compact and sober. The Lawson topology and the patch topology coincide (see [33]) and, with respect to the Lawson topology, X is a pospace. We will be interested in continuous domains that are coherent, that is, which have the property that the intersection of any 1.5. LAWSON-COMPACT CONTINUOUS DOMAINS 9 two Scott-compact saturated sets is Scott-compact. By the above, coherence implies stable local compactness for continuous dcpos. Proposition 1.9 ([15, Theorem III-5.8]) For a continuous domain X the following properties are equivalent: (1) X is Lawson-compact. (2) The Scott-compact saturated sets agree with the closed sets for the lower topology on X, that is, the co-compact topology agrees with the lower topology. (3) X is compact and coherent, that is, X with the Scott topology is stably compact. By the above, a Lawson-compact continuous domain becomes a compact pospace when endowed with the Lawson topology. Its Lawson-open upper sets are precisely the Scott-open sets and its Lawson-closed upper sets are precisely the Scott-compact saturated sets. In Section 4.3 we will apply Lemma 1.7 to reduce an order-convex Lawson-compact subset to its lower part, which is Scott-closed, and its upper part, which is compact saturated with respect to the Scott topology. Another important fact from [33] is Lemma 1.10 Every Scott-continuous retract of a Lawson-compact contin- uous domain is Lawson-compact. Most continuous domains that occur in semantics are coherent. Thus, it will not be a great restriction, if we restrict ourselves to Lawson-compact continuous domains in section 4.3. But there are exceptions. The following is an example of locally compact sober space which is not coherent. It is also an example of a continuous domain that is not Lawson-compact. Example We take a trivially ordered infinite set Y and attach two new elements a and b as minimal elements, that is we let a < y and b < y for each y ∈ Y , but a and b remain incomparable. This ordered set is a continuous domain, hence, locally compact and sober for the Scott topology, but it is not coherent: The subsets ↑a = {a} ∪ Y and ↑b = {b} ∪ Y are compact but their intersection Y is not. Chapter 2 Directed Complete Ordered Cones The concept of a directed complete ordered cone (d-cone, for short) will be introduced in this chapter. As these objects are not yet familiar in domain theory, we do so at a leisurely pace. We take some care in developing their properties, and we also study some classes of examples. The abstract probabilistic domains APD of Jones and Plotkin [24, 23] have influenced the development of the notion of a d-cone. These objects turn out to be the algebras of the monad given by the probabilistic pow- erdomain functor in the category of continuous domains with respect to a ‘convex structure’. Dealing with subprobabilities allows scalar multiplica- tion by real numbers between 0 and 1 only, addition is replaced by con- vex combinations. To overcome this inconvenience, Kirch introduced the extended probabilistic powerdomain and showed that this functor is still monadic and has continuous d-cones as algebras [29]. Although studying cones in a domain-theoretic setting is quite new, ordered cones have long played a role in various contexts. For ordered cones, it is natural to require addition, scalar multiplication and linear functionals to be monotone. D- cones can be seen as a variant of ordered cones: one requires the order to yield a dcpo and, accordingly, one requires addition, scalar multiplication and linear functionals to be Scott-continuous. Before we give detailed definitions we will name at least some previous 11 12 CHAPTER 2. DIRECTED COMPLETE ORDERED CONES occurrences of ordered cones. In [13] Fuchssteiner and Lusky studied them in a functional analytic setting. In Chapter 3 we will show that some of their results still hold for continuous d-cones. Other results about ordered cones, of which we will take advantage, are due to W. Roth [45]. He deals with ordered cones equipped with a quasiuniform structure proposed by Keimel and Roth in [27]. In the context of harmonic analysis and abstract potential theory, cones have been studied by Boboc, Bucur and Cornea [5]. Rauch has shown in [44] that a special class of their ordered cones, their standard H-cones, can also be viewed as continuous lattice-ordered d-cones with addition and scalar multiplication being Lawson continuous. Most of these cones — as is the case with d-cones — are not embeddable into real vector spaces as the cancellation law does not hold for addition. 2.1 D-Cones and Their Basic Properties We denote by R+ := {r ∈ R | r ≥ 0} the non-negative real numbers with their usual linear order and endowed with the Scott topology the only proper open sets of which are the intervals ]r, ∞[, r ∈ R+. Definition A set C is called a cone if it is endowed with two operations, an addition +: C × C → C and a scalar multiplication ·: R+ × C → C such that the following hold: there is a neutral element 0 ∈ C for addition making (C, +, 0) into a commutative monoid, that is, for all a, b, c ∈ C one has: (a + b) + c = a + (b + c) a + b = b + a a + 0 = a. Moreover, scalar multiplication acts on this monoid as on a vector space: 2.1. D-CONES AND THEIR BASIC PROPERTIES 13 for a, b ∈ C and r, s ∈ R+, one has 1 · a = a 0 · a = 0 (r · s) · a = r · (s · a) r · (a + b) = (r · a) + (r · b) (r + s) · a = (r · a) + (s · a). A function f : C → D between cones is called linear if, for all a, b ∈ C and r ∈ R+, one has f(a + b) = f(a) + f(b) f(r · a) = r · f(a). A cone C is an ordered cone if it is also endowed with a partial order ≤ such that addition and scalar multiplication considered as maps C × C → C and R+ × C → C, respectively, are order preserving in both variables. If the order is directed complete and if addition and scalar multiplication are Scott-continuous, then C is called a d-cone. Thus, a d-cone is at the same time a cone and a dcpo. In the case that C is a continuous domain, C is called a continuous d-cone. Note that we are using here the notions of Scott topology and continuity developed in Section 1.1 for conditionally directed complete partial orders; indeed it was precisely in order to define d-cones that we introduced these notions. The category of d-cones as objects and Scott-continuous linear maps as morphisms is denoted by CONE, and the full subcategory of continuous d-cones is called CCONE In the literature ordered cones are used in a slightly more general sense: For scalar multiplication one only requires x 7→ r·x: C → C to be monotone for every fixed r ≥ 0, whilst we require also that all the maps r 7→ r·x: R+ → C are order preserving. This stronger requirement implies that 0 is the least element, as 0 = 0 · x ≤ 1 · x = x for any x ∈ C. A d-cone also has a greatest element since the monotonicity of addition implies that the cone as a whole is directed and, hence, has a supremum which the is the greatest element. 14 CHAPTER 2. DIRECTED COMPLETE ORDERED CONES D-cones also have a topological flavour, but they are not necessarily topological cones: A topological cone is a cone C endowed with a topology such that both operations, addition (x, y) 7→ x + y : C × C → C and scalar multiplication (r, x) 7→ r·x: R+ ×C → C are jointly continuous. In contrast with classical analysis, we take R+ to have the Scott topology here. As noted in section 1.1, the Scott topology on a product of (conditionally) directed complete partial orders need not be the product of the Scott topologies on the factors, and so a Scott-continuous function defined on a product of (con- ditionally) directed complete partial orders need not be jointly continuous for the product of their Scott topologies. In particular, addition need not be jointly continuous on a d-cone. This phenomenon cannot occur if one of the two factors is a continuous (conditionally) directed complete partial order. Thus, scalar multiplication is jointly continuous on any d-cone, and addition is jointly continuous for continuous d-cones which, consequently, are topological cones for the Scott topology. We have discussed the relations between ordered cones, d-cones and topo- logical cones in some detail as we will apply results about topological cones and, especially, ordered cones to continuous d-cones. A simple example of a continuous d-cone is R+ := R+ ∪ {∞} with its usual linear order, addition and multiplication, extended to ∞ as follows: x + ∞ = ∞ = ∞ + x, x ∈ R+ x · ∞ = ∞, x ∈ R+ \{0} 0 · ∞ = 0. With this convention, addition and multiplication are Scott-continuous on R+. For any d-cone, scalar multiplication – which was supposed to be defined W↑ for r ∈ R+ only – can be extended to r = ∞ by defining ∞·x := {r·x|r ∈ R+}. The cone axioms will also hold for the extended scalar multiplication. It is straightforward to see that direct products of (continuous) d-cones are again (continuous) d-cones. Other examples are the extended proba- bilistic power domain, the space of lower semicontinuous functions and the dual d-cone. We postpone the definition and a more detailed discussion of these examples first examining some general properties of d-cones. 2.1. D-CONES AND THEIR BASIC PROPERTIES 15 2.1.1 The Way-Below Relation It is a useful property of d-cones that scalar multiplication preserves the way-below relation. We will see later that this is not true for addition, in general. Lemma 2.1 Let a, b be elements of a d-cone C with a b and let r ∈ R+. Then r · a r · b holds. Proof For r > 0 this follows from the fact that a 7→ ra is an order- isomorphism of C. If r = 0 then r · a = r · b = 0 is the least element of the d-cone and therefore compact. For some of our results we will need continuous d-cones where also ad- dition preserves the way-below relation. We give a name to this property: Definition The way-below relation on a d-cone is called additive, if a1 b1 and a2 b2 imply a1 + a2 b1 + b2. The additivity of the way-below relation is equivalent to the property that addition is an almost open map in the following sense: Proposition 2.2 Let C be a continuous d-cone. Then the way-below rela- tion is additive if and only if, for all Scott-open subsets U, V , the set ↑(U +V ) is Scott-open, too. Proof Suppose first that is additive. Let x ∈ ↑(U + V ). Then there are elements u ∈ U, v ∈ V such that u + v ≤ x. As C is continuous, there are elements u0 ∈ U, v0 ∈ V such that u0 u, v0 v. By the additivity of the way-below relation, u0 + v0 u + v ≤ x. This shows that ↑(U + V ) is Scott-open. For the converse, let u0 u and v0 v. Then u+v ∈ u0 + v0. As now the upper set generated by u0 + v0 is supposed to be Scott-open, there is an x in this set with x u + v. It follows that u0 + v0 ≤ x u + v. It will turn out that most of our examples of continuous d-cones have an additive way-below relation. 16 CHAPTER 2. DIRECTED COMPLETE ORDERED CONES Proposition 2.3 The way-below relation on R+ is additive. Proof On R+ the way-below relation is characterised by x y if and only if x < y or x = y = 0. It is straightforward that addition preserves this condition, and thus the way-below relation. The additivity of the way-below relation is preserved under products: Proposition 2.4 The way-below relation is additive on a product of con- tinuous d-cones with additive way-below relations. Q Proof The way-below relation on a product i∈I Xi of dcpos Xi with a smallest element ⊥i ∈ Xi can be characterised by the way-below relations i on Xi via (xi)i∈I (yi)i∈I if and only if there exist a finite subset E ⊆ I with xi = ⊥i for i 6∈ E and xi i yi for i ∈ E. The least element in a continuous d-cone is the neutral element 0. Thus, addition preserves the way-below relation in a product if this holds in each component. The way-below relation on the probabilistic powerdomain and on the cone of lower semicontinuous functions will be discussed later. There, we will also see an example of a continuous d-cone where the way-below relation is not additive. 2.1.2 Convex Sets On d-cones one has two notions of convexity: Definition A subset M of a cone C is called convex if a, b ∈ M implies r · a + (1 − r) · b ∈ M for all r ∈ [0, 1]. A subset M of a poset C is called order-convex if a, b ∈ M and a ≤ x ≤ b imply x ∈ M. A d-cone C is called locally convex if every point has a neighbourhood basis of Scott-open sets which are convex and order-convex. Principal filters ↑a and principal ideals ↓a are convex and order-convex for any a ∈ C, since scalar multiplication and addition on a d-cone are monotone. Together with the fact that the union of an increasing sequence 2.1. D-CONES AND THEIR BASIC PROPERTIES 17 of convex, order-convex sets is convex and order-convex, we see that a con- tinuous d-cone is always locally convex. This was pointed out to us by J.D. Lawson: Proposition 2.5 Every continuous d-cone C is locally convex. Indeed, ev- ery point in C has a neighborhood basis of Scott-open convex filters. Proof For a ∈ C let U be a Scott-open neighbourhood of a. Since C is continuous we can find a sequence (an)n∈N in U satisfying a1 a and S S an+1 an for all n ∈ N. Then V := an = ↑an is a Scott-open n∈N n∈N neighbourhood of a which is convex and order-convex and contained in U. In case the way-below relation is additive we can show even more: Lemma 2.6 For a continuous d-cone with an additive way-below relation, the Scott interior of any convex saturated set M is convex. Proof Let x, y ∈ int M and r ∈ [0, 1]. Then, there exist x0, y0 ∈ M with x0 x and y0 y. Using that the way-below relation is additive, we conclude r · x + (1 − r) · y r · x0 + (1 − r) · y0 ∈ M, as M is convex; hence, r · x + (1 − r) · y ∈ int M. There are other operations which preserve convexity. Lemma 2.7 Let M be a convex subset of a d-cone C. Then: 1. The Scott closure M is convex. 2. The saturation ↑M and the lower closure ↓M are convex. Proof For the first claim we use the formation of the Scott closure indicated before Lemma 1.4. In a first step we form the set M1 of all x ∈ C such that W↑ there is a directed family (ai) in M with x ≤ ai. The set M1 is convex. Indeed, for x, y ∈ M1 there are directed sets (ai) and (bj) in M such that W↑ W↑ x ≤ ai and y ≤ bj. For 0 ≤ r ≤ 1, the family rai + (1 − r)bj W↑ W↑ is also directed in M and rx + (1 − r)y ≤ r · ai + (1 − r) · bj = W↑ rai +(1−r)bj , whence rx+(1−r)y ∈ M1. We continue this procedure 18 CHAPTER 2. DIRECTED COMPLETE ORDERED CONES by transfinite induction defining convex sets Mn for ordinals n. (For limit S ordinals n we define Mn = m For nonempty subsets P and Q of any cone C and r ∈ R+, we may define r · P = {ra | a ∈ P } and P + Q = {a + b | a ∈ P, b ∈ Q} . Clearly, addition of subsets is associative, commutative, and the singleton zero set is a neutral element. Scalar multiplication satisfies all the cone axioms except that (r + s)P 6= rP + sP in general. Indeed, let C = R+ and P = {1, 2}, then 2P = {2, 4} but P + P = {2, 3, 4}, whence 2P 6= P + P . The situation changes, when we pass to convex subsets: Lemma 2.8 Let P,Q be subsets of a cone C and r ∈ R+. Then we have: 1. The convex hull of a scalar multiple is given by conv(r·P ) = r·conv P . 2. The convex hull of the sum is given by conv(P +Q) = conv P +conv Q. 3. If P,Q are convex, then r · P and P + Q are convex, too. 4. With the straightforward addition and scalar multiplication as defined above, the collection of all nonempty convex subsets of C is a cone. 5. If P and Q are convex, then the convex hull of the union is given by conv(P ∪ Q) = r · p + (1 − r) · q p ∈ P, q ∈ Q, r ∈ [0, 1] . The first and second statements of this lemma are straightforward and they imply the third statement. For the fourth statement the only notewor- thy part is the equality (r + s)P = rP + sP : Indeed, if r = s = 0, then the equation is trivial. If one of r and s is nonzero, then c ∈ r · P + s · P implies that there are elements a, b ∈ P 2.2. THE EXTENDED PROBABILISTIC POWERDOMAIN 19